[literacy] deep / shallow, local / global features in machine learning image processing

Shallow features ： The features extracted by shallow network are close to the input , Contains more pixel information , Mainly for some fine-grained information , For example, color. 、 texture 、 edge 、 Edge and angle information .

principle ： Shallow network receptive field is small , The overlapping area of receptive field is also small , So make sure the network captures more details

Deep features ： The features extracted by deep network are close to the output , Contains more abstract information , Semantic information , Mainly some coarse-grained information .

principle ： The receptive field increases , The overlapping area between receptive fields increases , Image information is compressed , What we get is some information about the integrity of the image .

Content based image retrieval (Content-based Image Retrieval, CBIR) Methods the features extracted from the image are used for retrieval .

Common image features mainly include color 、 Texture and shape , Including local features and global features .

      Local features are image descriptors extracted based on a certain region of the image , Such as scale invariant features SIFT(Scale Invariant Feature Transform).（ amount to CNN The shallow convolution part of the network ）; Local features usually come from the visual sensitive area of the picture .

     The global descriptor is based on the descriptor extracted from the whole image , Such as GIST. The compression rate of global features is high , But the distinction is not strong ;（ amount to CNN The deep convolution part of the network ）; Because a picture can only generate one global feature

Strong discrimination of local features , But too many , Therefore, various coding methods have been proposed , Such as BOF(Bag of Features, Feature bag ),Fisher vector (Fisher Vectors, FV), as well as VLAD (Vector of Locally Aggregated Descriptors) etc. .BOF,VLAD,FV Such descriptors usually inherit the partial invariance of local features , Such as translation 、 rotate 、 The zoom 、 Factors that are not semantically related, such as illumination and occlusion, remain unchanged .

Convolution layer characteristics and SIFT comparison

It has the following characteristics ：

（1） Convolution layer features are similar to dense SIFT features （ Through dense sampling in mesh format ）. Convolution layer characteristics and SIFT The same is a local feature , Corresponding to an area of the picture （ Can be CNN Each point on the feature map is inversely mapped back to the picture ）, Is a local feature .

（2） Convolution layer features are obtained by learning ,SIFT It's a manual type .CNN The convolution layer parameters can be optimized through iterative training for different data sets , And it can be further improved by simple modification （ Such as increasing the depth 、 Width etc. ） and SIFT The parameters of are fixed through pre-determined precision design .

（3） The characteristics of convolution layer are hierarchical . Different convolution layers have different semantic levels , For example, the characteristic map of shallow layer is usually some edges / Angle, etc , The middle layer is part of the object , The upper layer is usually a complete object . Choosing different layers may achieve completely different effects , So far, there is no optimal method to choose an optimal layer , It is usually achieved by testing the effect of multiple layers .SIFT Without using SP There is no hierarchy in the case of , It describes the edge / Low-level features, etc , That's why CCS take SIFT And CNN One of the reasons why integration works .

（4）CNN Convolution layer feature dimension ratio SIFT/SURF And other shallow features are much larger , And it's a lot of calculation , need GPU Assistance can achieve real-time effect , And because you have to store a lot of convolution feature maps , Space costs are also much greater . about PC As far as the opportunity is concerned , It's not a big problem , However, the future AI Will be possible everywhere ,CNN The use of mobile platforms will become a challenging problem . With the development of brain like computing , Various special chips for neural processing （ For example, it was developed by Chen Yunji of the Chinese Academy of Sciences DaDianNao, Google Recently developed TPU etc. ） Constantly emerging , This problem or will not be a problem .
————————————————
Copyright notice ： This paper is about CSDN Blogger 「Dust_Evc」 The original article of , follow CC 4.0 BY-SA Copyright agreement , For reprint, please attach the original source link and this statement .
Link to the original text ：https://blog.csdn.net/Dust_Evc/article/details/123854535